1 /* NOLINT(build/header_guard) */
2 /* Copyright 2013 Google Inc. All Rights Reserved.
3 
4    Distributed under MIT license.
5    See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
6 */
7 
8 /* template parameters: EXPORT_FN, FN */
9 
EXPORT_FN(CreateBackwardReferences)10 static BROTLI_NOINLINE void EXPORT_FN(CreateBackwardReferences)(
11     size_t num_bytes, size_t position,
12     const uint8_t* ringbuffer, size_t ringbuffer_mask,
13     ContextLut literal_context_lut, const BrotliEncoderParams* params,
14     Hasher* hasher, int* dist_cache, size_t* last_insert_len,
15     Command* commands, size_t* num_commands, size_t* num_literals) {
16   HASHER()* privat = &hasher->privat.FN(_);
17   /* Set maximum distance, see section 9.1. of the spec. */
18   const size_t max_backward_limit = BROTLI_MAX_BACKWARD_LIMIT(params->lgwin);
19   const size_t position_offset = params->stream_offset;
20 
21   const Command* const orig_commands = commands;
22   size_t insert_length = *last_insert_len;
23   const size_t pos_end = position + num_bytes;
24   const size_t store_end = num_bytes >= FN(StoreLookahead)() ?
25       position + num_bytes - FN(StoreLookahead)() + 1 : position;
26 
27   /* For speed up heuristics for random data. */
28   const size_t random_heuristics_window_size =
29       LiteralSpreeLengthForSparseSearch(params);
30   size_t apply_random_heuristics = position + random_heuristics_window_size;
31   const size_t gap = 0;
32 
33   /* Minimum score to accept a backward reference. */
34   const score_t kMinScore = BROTLI_SCORE_BASE + 100;
35 
36   BROTLI_UNUSED(literal_context_lut);
37 
38   FN(PrepareDistanceCache)(privat, dist_cache);
39 
40   while (position + FN(HashTypeLength)() < pos_end) {
41     size_t max_length = pos_end - position;
42     size_t max_distance = BROTLI_MIN(size_t, position, max_backward_limit);
43     size_t dictionary_start = BROTLI_MIN(size_t,
44         position + position_offset, max_backward_limit);
45     HasherSearchResult sr;
46     sr.len = 0;
47     sr.len_code_delta = 0;
48     sr.distance = 0;
49     sr.score = kMinScore;
50     FN(FindLongestMatch)(privat, &params->dictionary,
51         ringbuffer, ringbuffer_mask, dist_cache, position, max_length,
52         max_distance, dictionary_start + gap, params->dist.max_distance, &sr);
53     if (sr.score > kMinScore) {
54       /* Found a match. Let's look for something even better ahead. */
55       int delayed_backward_references_in_row = 0;
56       --max_length;
57       for (;; --max_length) {
58         const score_t cost_diff_lazy = 175;
59         HasherSearchResult sr2;
60         sr2.len = params->quality < MIN_QUALITY_FOR_EXTENSIVE_REFERENCE_SEARCH ?
61             BROTLI_MIN(size_t, sr.len - 1, max_length) : 0;
62         sr2.len_code_delta = 0;
63         sr2.distance = 0;
64         sr2.score = kMinScore;
65         max_distance = BROTLI_MIN(size_t, position + 1, max_backward_limit);
66         dictionary_start = BROTLI_MIN(size_t,
67             position + 1 + position_offset, max_backward_limit);
68         FN(FindLongestMatch)(privat,
69             &params->dictionary,
70             ringbuffer, ringbuffer_mask, dist_cache, position + 1, max_length,
71             max_distance, dictionary_start + gap, params->dist.max_distance,
72             &sr2);
73         if (sr2.score >= sr.score + cost_diff_lazy) {
74           /* Ok, let's just write one byte for now and start a match from the
75              next byte. */
76           ++position;
77           ++insert_length;
78           sr = sr2;
79           if (++delayed_backward_references_in_row < 4 &&
80               position + FN(HashTypeLength)() < pos_end) {
81             continue;
82           }
83         }
84         break;
85       }
86       apply_random_heuristics =
87           position + 2 * sr.len + random_heuristics_window_size;
88       dictionary_start = BROTLI_MIN(size_t,
89           position + position_offset, max_backward_limit);
90       {
91         /* The first 16 codes are special short-codes,
92            and the minimum offset is 1. */
93         size_t distance_code = ComputeDistanceCode(
94             sr.distance, dictionary_start + gap, dist_cache);
95         if ((sr.distance <= (dictionary_start + gap)) && distance_code > 0) {
96           dist_cache[3] = dist_cache[2];
97           dist_cache[2] = dist_cache[1];
98           dist_cache[1] = dist_cache[0];
99           dist_cache[0] = (int)sr.distance;
100           FN(PrepareDistanceCache)(privat, dist_cache);
101         }
102         InitCommand(commands++, &params->dist, insert_length,
103             sr.len, sr.len_code_delta, distance_code);
104       }
105       *num_literals += insert_length;
106       insert_length = 0;
107       /* Put the hash keys into the table, if there are enough bytes left.
108          Depending on the hasher implementation, it can push all positions
109          in the given range or only a subset of them.
110          Avoid hash poisoning with RLE data. */
111       {
112         size_t range_start = position + 2;
113         size_t range_end = BROTLI_MIN(size_t, position + sr.len, store_end);
114         if (sr.distance < (sr.len >> 2)) {
115           range_start = BROTLI_MIN(size_t, range_end, BROTLI_MAX(size_t,
116               range_start, position + sr.len - (sr.distance << 2)));
117         }
118         FN(StoreRange)(privat, ringbuffer, ringbuffer_mask, range_start,
119                        range_end);
120       }
121       position += sr.len;
122     } else {
123       ++insert_length;
124       ++position;
125       /* If we have not seen matches for a long time, we can skip some
126          match lookups. Unsuccessful match lookups are very very expensive
127          and this kind of a heuristic speeds up compression quite
128          a lot. */
129       if (position > apply_random_heuristics) {
130         /* Going through uncompressible data, jump. */
131         if (position >
132             apply_random_heuristics + 4 * random_heuristics_window_size) {
133           /* It is quite a long time since we saw a copy, so we assume
134              that this data is not compressible, and store hashes less
135              often. Hashes of non compressible data are less likely to
136              turn out to be useful in the future, too, so we store less of
137              them to not to flood out the hash table of good compressible
138              data. */
139           const size_t kMargin =
140               BROTLI_MAX(size_t, FN(StoreLookahead)() - 1, 4);
141           size_t pos_jump =
142               BROTLI_MIN(size_t, position + 16, pos_end - kMargin);
143           for (; position < pos_jump; position += 4) {
144             FN(Store)(privat, ringbuffer, ringbuffer_mask, position);
145             insert_length += 4;
146           }
147         } else {
148           const size_t kMargin =
149               BROTLI_MAX(size_t, FN(StoreLookahead)() - 1, 2);
150           size_t pos_jump =
151               BROTLI_MIN(size_t, position + 8, pos_end - kMargin);
152           for (; position < pos_jump; position += 2) {
153             FN(Store)(privat, ringbuffer, ringbuffer_mask, position);
154             insert_length += 2;
155           }
156         }
157       }
158     }
159   }
160   insert_length += pos_end - position;
161   *last_insert_len = insert_length;
162   *num_commands += (size_t)(commands - orig_commands);
163 }
164